AG-Pose: Optimized for Qualcomm Devices

AG-Pose is a category-level 6D object pose estimation model that predicts rotation, translation, and size of objects from RGB-D input. It uses DINOv2-S (ViT-S/14) as the RGB backbone and PointNet++ for 3D point cloud processing, with an Instance-Adaptive Keypoint Detector (IAKD) and Geometric-Aware Feature Aggregator (GAFA) for robust pose estimation across 6 object categories.

This is based on the implementation of AG-Pose found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.50, ONNX Runtime 1.30.0	Download
QNN_DLC	float	Universal	QAIRT 2.50	Download

For more device-specific assets and performance metrics, visit AG-Pose on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

• Custom weights (e.g., fine-tuned checkpoints)
• Custom input shapes
• Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for AG-Pose on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.pose_estimation

Model Stats:

• Model checkpoint: epoch_30.pt
• Input resolution: 1 x 3 x 224 x 224
• Number of parameters: 30.4M
• Model size (float): ~116 MB
• Supported categories: bottle, bowl, camera, can, laptop, mug

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
backbone	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	50.344 ms	1 - 351 MB	NPU
backbone	ONNX	float	Qualcomm® Dragonwing™ Q-8750	50.344 ms	1 - 351 MB	NPU
backbone	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	52.827 ms	1 - 403 MB	NPU
backbone	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	52.827 ms	1 - 403 MB	NPU
rotation_head	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	0.16 ms	0 - 25 MB	NPU
rotation_head	ONNX	float	Qualcomm® Dragonwing™ Q-8750	0.16 ms	0 - 25 MB	NPU
rotation_head	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	0.158 ms	0 - 24 MB	NPU
rotation_head	QNN_DLC	float	Qualcomm® Dragonwing™ Q-8750	0.158 ms	0 - 24 MB	NPU

License

• The license for the original implementation of AG-Pose can be found here.

References

• AG-Pose: Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.